Flexible and robust deep learning models for integrative analysis of single-cell RNA sequencing data

单细胞RNA测序（scRNA-Seq）允许单独测量细胞，这使我们能够研究细胞与细胞的异质性，发现新的细胞群并了解它们的关系。ScRNA-Seq正在迅速提高我们对生物细胞功能多样性的理解。最近的研究报道了大规模scRNA-Seq数据集。因此，为了全面分析复杂的生物过程，需要整合来自多个实验的不同类型的单细胞RNA数据，这需要灵活但严格的计算框架。在这个研究项目中，我和我的团队将致力于开发灵活而强大的深度学习模型，以整合来自不同物种和实验的scRNA-Seq数据，探索细胞异质性。这包括三个研究主题：用于整合scRNA的深度张量因子分解方法-来自人类和小鼠的Seq数据以分层细胞类型。许多测序实验已经产生了大量来自人和小鼠的单细胞的标记和未标记数据集。我们将开发基于深度张量因子分解的方法，将张量因子分解与自我监督深度学习方法相结合，以利用这些不同的数据集来研究细胞类型异质性。我们将研究数据集之间独特的细胞关系。具有多主题单细胞表达的大块组织细胞类型去卷积的迁移学习方法。了解相关组织中的细胞类型组成是迈向组织表征数字化方法的重要一步。我们将开发半监督迁移学习方法，利用scRNA-Seq数据中的细胞类型特异性基因表达来表征复杂组织中大量RNA-Seq数据中的细胞类型组成。这可以使细胞类型特异性基因表达信息从一个数据集转移到另一个数据集，并表征复杂组织的细胞异质性。生物学上可解释的深度学习模型，用于表征细胞类型特异性调控网络。我们的目标是开发一个基于注意力的图卷积网络（GCN），用于使用大型单细胞转录组学数据集对细胞类型特异性调控网络进行建模。通过该模型，我们可以通过大型调控网络将基因表达与细胞表型联系起来。我们将验证模型的可解释性模拟数据和生物系统与已知的地面真相，并揭示调控蛋白在探索系统。通过这项研究计划，我们将通过开源软件工具提供灵活而强大的深度学习模型。通过解决单细胞转录组中的几个主要挑战，我们的新模型和工具将有助于释放scRNA-Seq技术在研究细胞异质性方面的全部潜力，这也将为加拿大的大型动物科学，农业和植物科学带来重大利益。

Single cell RNA sequencing (scRNA--Seq) allows for cells to be measured individually, which permits us to investigate cell- to -cell heterogeneity, discover novel cell populations and understand their relationships. ScRNA--Seq is rapidly improving our understanding of functional diversity among biological cells. Recent studies reported large-scale scRNA--Seq data sets. Hence, to comprehensively analyze complex biological processes, different types of single -cell RNA data from multiple experiments need to be integrated, which require flexible but rigorous computational frameworks. In this research program, my team and I will work toward development of flexible and robust deep learning models to integrate scRNA--Seq data from different species and experiments to explore cellular heterogeneity. This includes three research themes: Deep tensor factorization approaches for integration of scRNA--Seq data from human and mouse to stratify cell types. Many sequencing experiments have generated large number of labeled and unlabeled datasets of single cells from human and mouse. We will develop deep tensor factorization-based approaches, which integrate tensor factorization with self-supervised deep learning methods, to leverage these diverse datasets for investigating cell type heterogeneity. We will study unique cellular relationships between the datasets. Transfer learning approaches for bulk tissue cell type deconvolution with multi-subject single--cell expression. Understanding of cell type composition in relevant tissues is an important step toward a digital approach to tissue characterization. We will develop semi--supervised transfer learning methods to utilize cell type-specific gene expression from scRNA--Seq data to characterize cell type compositions from bulk RNA--Seq data in complex tissues. This can enable the transfer of cell type specific gene expression information from one dataset to another and the characterization of cellular heterogeneity of complex tissues. Biologically interpretable deep learning models for characterizing cell type-specific regulatory networks. We aim to develop an attention--based graph convolutional network (GCN) for modeling cell type-specific regulatory networks using large single -cell transcriptomics datasets. Through the model we can connect gene expression to cellular phenotypes through large regulatory networks. We will validate the model's interpretability on simulated data and biological systems with known ground truth, and reveal regulatory proteins in underexplored systems. Through this research program, we will deliver our flexible and robust deep learning models through open-source software tools. By addressing several major challenges in single cell transcriptome, our new models and tools will help unleash the full potential of scRNA--Seq technologies for studying cellular heterogeneity, which will also bring significant benefits to large animal science, agriculture and plant science in Canada.